Anthropic 发布了名为 Inkling 2 Mini 的小型开源权重大语言模型,该模型采用独特的“循环 Transformer”架构,通过重复运行中间层来提升推理能力而无需增加参数量 1。Inkling 2 Mini 的总参数为 26.4B,激活参数为 3.4B;其核心设计是将模型 24 层中的中间 12 层对每个 Token 重复运行 3 次 1。
在部署规格方面,该模型权重原生采用 NVFP4 精度(最后四层为 MXFP8),内存占用约为 20GB,支持在 MacBook Pro 等设备上离线运行 1。尽管体积小巧,Inkling 2 Mini 在多项基准测试中超越了同尺寸的其他开放权重模型,具体得分为 DeepSWE v1.1 的 48.2% 和 Vals Index v2.2 的 46.5% 1。目前,该模型已通过 Tinker、Hugging Face 和 Ollama 等平台提供推理及微调服务 1。
研究实验还揭示了该架构对链式思维(CoT)监控的影响:随着循环次数从 1 次增加到 3 次,模型的无 CoT 时间视界从 16 秒提升至 37 秒 1。观察显示,更多的循环次数使得 CoT 变短但仍保持可读性,且未出现明显的监控逃避趋势 1。
Anthropic has introduced Inkling 2 Mini, a small open-weight large language model featuring a unique "recurrent transformer" architecture 1. The model comprises 26.4 billion total parameters, with 3.4 billion active parameters per inference step 1. Its core design repeats the middle 12 layers out of a total of 24 for every token processed, enhancing reasoning capabilities without increasing the parameter count 1.
Despite its architectural complexity, Inkling 2 Mini is optimized for local deployment with a memory footprint of approximately 20GB 1. The model weights are natively stored in NVFP4 precision, except for the final four layers which use MXFP8 format 1. This efficiency enables offline operation on devices such as a MacBook Pro 1. The model is available for inference and fine-tuning via platforms including Tinker, Hugging Face, and Ollama 1.
In benchmark evaluations, Inkling 2 Mini surpassed other open-weight models of similar size, particularly in agentic coding tasks 1. It achieved a score of 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2 1. Research into the recurrent structure showed that increasing repetitions from one to three extended the no-chain-of-thought (CoT) time horizon from 16 seconds to 37 seconds 1. Additionally, more repetitions resulted in shorter yet readable CoT outputs, with no significant trends of monitoring evasion observed during explainability experiments 1.
评论
还没有评论,欢迎留下第一条。